HNHacker News
TopNewBestAskShowJobs

somethingsome

462 karma · joined June 13, 2020

Contact me at somethingsomehn@protonmail.com
submissionscomments
somethingsome··on Flux better than Stable Diffusion
On another note, after posting, I went to the main website and they display several user examples. They promise 'text' in images but everything I saw failed miserably, so I guess we are not there yet.

For the metric to evaluate this kind of network, to me it's impossible to define it, or at least correctly, it's too subjective, if you put a metric on it, it will influence the results and therefore limit the images you can create. (not even speaking about the kind of images we want it to create)

So maybe the best we can have is a very generic metric for training, and finetuning it to some style afterward depending on your usage.

So I don't really trust 'objective ranking', but we could take all those networks, fine tune them on some well defined style and evaluate them with ground truth images when we ask to make the original piece, it would at least be some kind of proxy on the finetuning-ability of the network.

I'm skipping a lot of details such as: how to generate the prompt corresponding to the desired output, but without thinking too much about it, it could be like finding some fix-point function, and there, we can create a metric. (but this fixed function would depend on the model, and probably we can't use it to evaluate other models, and having a fixed function for all models would not compare fairly)

Note: thinking out loud

somethingsome··on Flux better than Stable Diffusion
Hum, I actually submitted it so I can keep track of it (saved in my submitted record on HN) , it seemed promising, I hope to run some tests soon.

Didn't saw the organic marketing, I'll check better next time ;)

I just discovered it from https://youtu.be/stOiAuyVnyQ?si=eOVVVcUxJMClYZFa

somethingsome··on Floor796
If you subscribe you receive a mail with the new added cell automatically ;)
somethingsome··on Dragonfly: An optical telescope built from an array of off-the-shelf Canon lens
Is it possible to download somewhere in full quality a raw image captured with it?
somethingsome··on Neo (Keyboard Layout)
I'm a user of Bepo (French)[1] for many years and I'll never look back to classical azerty and qwerty layouts.

Bepo doesn't split the keyboard per 'usage' per se, but remains extremely fast to use for programming and writing English or German, the German letters are logically set under the corresponding letters in bepo with a modification key.

[1] https://bepo.fr/wiki/Accueil

somethingsome··on The Windows 11 Start Menu is comically bad
For me at least (win 10) the win key + search randomely but systematically after some time fails to show me anything. So I'm stuck in an empty menu forever. I have no idea what causes this, I'm a heavy user and this pattern appears on all my machines.. My only solution is to forcefully kill the Explorer process, if that fail, I need to restart the whole computer..

Also, 'win + Calc' takes ages to 'find the app, launch the app'.

Finally, when I write the same string of characters to launch a program, sometimes it launches another one with a similar name.

Eg: visual studio and visual studio code, or randomely it open edge or chrome and does a Google search..

I suspect part of those issues are there because I type very fast.

Honestly that shouldn't be so bad. My computers are mostly i9 with plenty of ram.

somethingsome··on What happens when we pay public high school teachers based on performance
One of the problem is that you can train the students to overfit any test, and that's worse than cheating which represent only a small number of students usually.

So by incentiving the teachers to make the students perform well in some tests, why teach real lessons when you can just throw at them cheat codes for the test.

somethingsome··on Ask HN: Why are we hanging on to anachronistic academic paper formats
I don't know, something around 1 to 3 per day for the last 7 years maybe, I read and write papers as a living.. I never really had any problem.. I mostly have issues with epub like formats..

Of course tts is a different beast, but.. Try to read science in general with tts.. It doesn't make sense.. And you need time to process each sentence and put it into context.. So even with reflow, tts would be a nightmare for science content.

somethingsome··on Vala Programming Language
While sciter seems nice, their

> In almost 10 years, Sciter UI engine has become the secret weapon of success for some of the most prominent antivirus products on the market: Norton Antivirus and Internet Security, Comodo Internet Security, ESET Antivirus, BitDefender Antivirus, and others. The use of HTML/CSS has allowed their UI to stay in touch with modern GUI trends throughout all these years, and will continue to well into the future.

Made me laugh a bit, I usually find antivirus softwares out of touch with good design and really impractical due to that.

somethingsome··on Ask HN: Why are we hanging on to anachronistic academic paper formats
I read most of my papers on my phone without any issues, I have absolutely no idea what you are talking about.

With Adobe PDF, you double click on one column, it zooms onto it perfectly, you double click again, it goes full page.. Navigation is super easy, the zoom feature perfectly fit the column to my screen.

somethingsome··on Ask HN: Why are we hanging on to anachronistic academic paper formats
2 column is so helpful in reading the paper imo. It's easier to jump and see previous parts, no need for page width equations that use a lot of empty space. And basically the style is enjoyable for me, I really don't see what is not approcheable.

Having a different style of the paper will not magically make it easier to understand. Focusing too much on the format instead of the content is counter productive.. And having somewhat standardized sections allows an academic to jump to juicy parts with enough experience.

somethingsome··on Ask HN: Why are we hanging on to anachronistic academic paper formats
You don't like it.. I know many academics, and I'm one of them that like very much 2 column information dense papers..
somethingsome··on CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
You use SfM to find the first point cloud. However SfM is based on the hypothesis that the same point 'moves' linearly in between any two views. This hypothesis is important because it allows you to match a point in two pictures, and given the distance between the two images, you can triangulate the point in space. Therefore find it's depth.

However, non-Lambertian points move non linearly in viewing space (eg a specular point depends on the viewer pose).

So, automatically, their positions in space will be false, and you'll have floating points.

Gaussian 'splats' may have the potential to render non-Lambertian stuff using for example the spherical harmonics (even if I don't think the viewer use them if I'm not mistaken). But, capturing non-Lambertian points is very difficult and an open research problem.

somethingsome··on CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
This technique exists from more than 10 years, and real time renderers exist too from very long.
somethingsome··on CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
3DGS is absolutely not good with non Lambertian materials..

After testing it, if fails in very basic cases. And it is normal that it fails, non Lambertian materials are not reconstructed correctly with SfM methods.

somethingsome··on CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
For NeRF this problems exists. However, in the past it was already solved for gaussian splatting. Usually you define a normal field over the (2D) splat, This allows you to have phong shading at least.

It is not too difficult to go to a 2D normal field over the 3D gaussians..

somethingsome··on CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
I know, I don't get the fuzz either, I've coded real-time gaussian splat renderers >7 years ago with LOD and they were able to show any kind of point cloud.

They worked with a basic 970 GTX on a big 3d screen and also on oculus dk2.

somethingsome··on CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
Of course! And this was done many times in the past, probably with better results than current deep learning based gaussian splatting where they use way too many splats to render a scene.

Basically the problem with sparse pictures and point clouds in general is their lack of topology and not precise spatial position. But when you already have the topology (eg with a mesh), you can extract (optimally) a set of points and compute the radius of the splats such that there are no holes in the final image (and their color). That is usually done with the curvature and the normal.

The 'optimally' part is difficult, an easier and faster approach is just to do a greedy pass to select good enough splats.

somethingsome··on From scratch OpenGL and shaders with raw Xlib
It is maintained by the w3c, not used extensively by the businesses that will employ the students, it's easy to go from OpenGL to Vulkan. I can reconsider it and happily use it if I have more arguments in favor of it, but from what I see it doesn't seem groundbreaking and adds a layer of complexity in comparison to OpenGL 4.x without really improving the learning curve or the capabilities.
somethingsome··on From scratch OpenGL and shaders with raw Xlib
I was teaching OpenGL for several years, we did C++ and OpenGL and then, during covid, we switched to web solutions.. What a pain.

Not the WebGPU or WebGL per se, but teaching engineers not familiar with Javascript or web development, local servers, CORS, etc.. adds a whole new level of difficulty.

somethingsome··on Google suspends romance author's account for writing sexually explicit content
I have so many problems with nextcloud! It kills itself from time to time, often during updates. Randomly problems with folder rights when everything was working before. It's slow and the code is php. The official documentation and forum are a mess.

I have a lot of files, the database is not able to handle them correctly, even the cleaning commands take hours and do not work.. I regularly lose files in the database and need to re-upload them from the file in the server..

The app on Android has a lot of issues synchronizing my files, too many too often I guess.

I'm hoping to see a good alternative to nextcloud one day.

somethingsome··on Why isn't preprint review being adopted?
A fun system would be: you 'have to' peer review publicly a (part of a) paper (at least) when you cite it in your paper.

So that often cited paper get a lot of different (small) public reviews that can be curated from time to time, obscure papers get at least one review justifying why it's relevant to cite them in the new work.

Some could argue that this is too much work added to the writing process.. But.. At the same time.. Shouldn't we read the papers we cite? Why not automatically write a small review of it? It has not to be huge, only the justification on why we (can) use it in our work.

somethingsome··on Why isn't preprint review being adopted?
IMO 'public trust in science is at an all time low' is because of bad journalism more than bad papers.

Papers are written by academy-type individuals for academy-type individuals, not for consumption by non experts. An academic is usually pretty fast to determine if a paper is to be trusted.

So interpreting and extrapolating to the extreme the results of a minor paper in an obscure journal is more bad journalism than bad science.

Then we wonder why people don't trust science..

somethingsome··on Show HN: Lapdev, a new open-source remote dev environment management software
Main current main pain point with devcontainers is to run a gui app remote, wathever I do, the gui opens only on the server. I'm wondering if this solution can export gui remotely?
somethingsome··on Google Scholar PDF Reader
Mostly too slow for a lot of content, not every content is supported, not easy to keep it open at the right page, no comments, not easy to find the right tab, etc.
somethingsome··on Google Scholar PDF Reader
Just so you know: normally it scroll down so that the reference is on top of the page.

But most importantly.. ALT+'left arrow' allows you to go back before you clicked on the citation! It doesn't work all the time, but usually it does after some left arrows ;)

Also, in Android: you can click on the 'scrolling sign' on the right of the pdf and specify the page, or see the link to 'jump back' to before you clicked on a link!

I hope that will help

somethingsome··on Hi everyone yes, I left OpenAI yesterday
It's actually both!

A teacher can usually adapt the content depending on its audience, I would not teach the research in my field at the same level to professionals, PhDs, master students, bachelor students, amateurs, or even school students.

If what I'm teaching is fairly complex, it requires a lot of background that I could teach, but I would not have the time to do so, because it would be to the detriment of other students. So, while I usually teach 'from scratch', depending on my audience I will obfuscate some details (that I can answer separately if a question is asked) and usually I will dramatically change the speed of the lessons depending on the previous background, because I need to assume that the student has the prerequisite background to understand at that speed fairly complex material.

As an example, I gave some explanations to a student from zero to transformers, it took several hours with lots of questions, the same presentation to a teacher not in the field took me 1h30 and to a PhD in a related field took 25 minutes, the content was exactly the same, and it was from scratch, but the background in the audience was fairly different.

somethingsome··on Bayesian Analysis with Python
Self plug for stan on windows in a docker dev environment:

https://github.com/dbonattoj/cookiecutter-pystan

somethingsome··on What Makes for 'Good' Mathematics?
> I think the best and most beautiful mathematics is driven directly by applications, and made better for the pressure to apply to a real problem (not a problem that just mathematicians care about).

That's the thing, in my opinion, when you study math, you start with applications, you work your way to solve complex applicative problems, and then, those problems start to be less interesting, in part because they become easy for you, and the other part is that most of the problem is just 'mechanical and technical' work. Not really interesting.

But what you like in math, it's not the maths themselves, it's to be challenged by problems. So you start to work on abstracts problems, until they become easy, then you go one more step in the abstraction, to find challenging problems, etc..

So basically, I find it somewhat natural to be driven always to more abstract theory, with maybe the hope that one day they will be applied. It's far from you to solve too 'technical' problems in the sense that 'it's just an application of XYZ problem to solve, now, I don't want to implement it, as it will take me a lot of time, while I can find pleasure again in discovering new hard things'

I'm also on the applicative side, but I understand the attraction of the abstract side, and sometimes, I allow myself to work with just a bit of higher abstraction.

So with enough luck, some day, a mathematician, with the right level of abstraction at that moment, will discover an applicative problem that is of your interest, and dedicate to it some time, maybe a lot of time, and maybe even, his/her whole carreer, making it more abstract, elegant and practical.

somethingsome··on LoRA from scratch: implementation for LLM finetuning
Thanks, I'll meditate on that and re read the paper with this view in mind.

The last sentence makes sense to me, if the finetuning job changes significatively more the weights of other layers than just the last one, it is kinda normal to to use Lora on them. I had the impression that it was rarely the case, but I must be mistaken. I'll think about applications where this is the case.

← PreviousPage 7 of 9Next →